Image Retrieval Algorithm Based On Deep Convolutional Network And Hash Coding

In recent years,resources of image data that people can get from the Internet have increased dramatically.In order to search the images which people actual needs from a large number of data quickly and effectively,image retrieval technology has received considerable attention.Due to its encouraging efficiency in both speed and storage,hashing is applied more and more by researchers to the similarity search problem of multimedia data such as images.The traditional hash algorithm usually starts with the process of extracting hand-crafted visual features.Then,the extracted feature vectors are quantized to generate a series of binary code.At present,the popularity of Convolutional Neural Network makes researchers has led researchers to solve problems in image processing fields.Considering the superiority of Convolutional Neural Network in feature extraction,we combine it with hash algorithm for image retrieval.The main work of this article is as follows:(1)This paper proposes a deep hash algorithm of image retrieval based on multi task learning.On the one hand,image classification information and similarity information between images are used for training the model simultaneously,so as to maintain the semantic similarity between hash codes.And we use the Spatial Pyramid Pooling to extract more accurate features.At the same time,this paper puts forward two kinds of training strategies,which are combined and separated,combined with image classification loss and contrast loss in different ways.Experimental results on two datasets CIFAR-10 and NUS-WIDE show that our algorithm has better retrieval results compared with other hashing algorithms.(2)Aiming at the limitation of Spatial Pyramid Pooling,a Multi Scale Fusion Pooling method(MSFP)is proposed in this paper,which could fuse multiple scale information in the image,and effectively reduce the number of parameters in the network model.On the two datasets of CIFAR-10 and NUS-WIDE,we compare the deep hash algorithms of image retrieval based on these two pooling methods.The experimental results show that the Multi Scale Fusion Pooling method proposed in this paper could effectively reduce the computational complexity of the model training stage as well as the storage space taken up by the final training model,and also maintain good retrieval performance.